Serverless promises huge cost reduction in operations, increase of stability and scalability and overall increase of developer and user happiness. We bring the serverless delta architecture to kubernetes using open source standards and technologies including Apache Iceberg, Knative/Kubernetes, Cloud Object Store, Apache Kafka and KFServing into a single open source reference architecture capable of running on any Kubernetes service in the Cloud (IBM, Amazon, Google, Microsoft, …) as well as on on-premises datacenters (RedHat Openshift, …). The architecture combines batch and stream processing, data lakes, data catalogs and data warehouses, business intelligence and machine learning.
Why are most DeepRL publications on playing video games? Because we have a perfect simulator? Can’t we use the abundance of historic financial data as a simulator for trading bots. I meant agents of course. This workshop covers the basics of DeepRL by explaining how AlphaGo Zero works and takes you towards an end to end implementation of a DeepRL based bot for trading. This is my contribution to democratize Artificial Intelligence in Finance.
Using the Elyra pipeline editor for JupyterLab and VSCode we've developed a no-code/low-code environment which allows for mixing and matching ready-made components (CLAIMED component library), python/R scripts and notebooks as well as KubeFlow pipeline components into full round trip no-code/low-code editing with support for local execution as well as on Airflow or KubeFlow on top of Kubernetes. CLAIMED supports scikit-learn, pandas, R, TensorFlow, PyTorch and ApacheSpark and can be easily extended to support anything which can run in a (docker) container. For TrustedAI, bias detection/fairness, explainability and adversarial robustness are supported. This topic has been peer-reviewed and published => http://conference.scipy.org/proceedings/scipy2021/pdfs/ivan_nesic.pdf
Alexa, Cortana, Siri and Google Assistant have one thing in common. They send your data unencrypted to long term storage and analysis to the cloud. MyCroft tries to tackle this problem by providing a fully open source personal assistant framework running on custom hardware, Raspberry Pi or your Desktop. In this talk we demo MyCroft and assess it's architecture from a privacy point of view. (Optionally: we show you how custom MyCroft skills can easily be created)
Cloud data science platforms are a nice thing as they completely remove the burden to operate them and let employees concentrate on solving problems. But this also often means complete vendor lock-in. In this tutorial we present the Open Source Data Science Platform which runs on any Kubernetes environment. In the cloud, in your data center and even on your laptop.
In this tutorial you will learn how to setup and use the Open Source Data Science Platform based on Jupyter-Lab, Elyra-AI, Apache Spark, Apache Superset, Seldon, Kubernetes, Kubeflow Pipelines and S3.
Although an instance of the Open Source Data Science Platform is provided to you, we'll briefly walk you through installation and configuration. Although a laptop is sufficient to run, we recommend to use a server with at least 16 GB of main memory and 8 cores.
Central to the platform is a data lake. Here S3 is used as it emerged to a standard across all major cloud providers and also through Ceph and Minio open source alternatives are available for on-prem usage. In this tutorial we will use minio as a foundation to our data lake.
Apache Spark will be used as engine for data ingestion as well as providing a JDBC access layer to S3 data to be consumed by Apache Superset, the open source BI tool. Having a BI tool allowing for arbitrary stakeholders to explore data without coding skills is a key catalyst for getting buy in from stakeholders to promote data science projects. At the same time, the ODBC/JDBC interface allows for existing, commercial BI tools to connect to the data lake as well.
For programmers and data scientists we show how to create an operator library using jupyter notebooks and python scripts which can be dragged and dropped to the pipeline canvas of Elyra-AI and automatically compiled and deployed to Kubeflow Pipelines pushing all your workload to Kubernetes.
Finally, with ML Ops, which includes pushing ML models to Seldon for seamless model serving on Kubernetes and assessing and monitoring model quality (performance, bias, fairness, adversarial robustness, drift, ...).
Data privacy is a huge concern and often prevents ML and AI project from flourishing.
In this talk we’ll introduce you to federated learning and homomorphic encryption. After we’ve covered the theoretical aspects we’ll see how they can be used in practice.
We conclude with an outlook on the future of these technologies.
Trusted AI – Building Reproducible, Unbiased and Robust AI Pipelines using the python OpenSource stack
We are just in the middle of the DeepLearning hype. A lot of things are done, but enterprises are still struggling on production deployments. One of the reasons is that untrusted AI raises a lot of concerns. In this talk we’ll explain data lineage, bias detection, adversarial robustness and model explainability and how it can be achieved using an open source stack. Key takeaways:
- Introduction to the key concerns on large scale AI adoption in the Enterprise
- Explaination of the terms data lineage, bias detection, adversarial robustness and model explainability
- Presentation of an open source framework to migitage these aspects
In 2019 we've noticed a tremendous shift of client demand on model creation to model deployment and monitoring (CI/CD). This indicates a further step towards maturity of wider AI adoption within enterprises. In this talk we'll introduce you to the latest developments in the most widely used DeepLearning framework TensorFlow 2.x. We show how Kubernetes and Kubeflow Pipelines work and how Open Data Hub provides a Open Source powered platform for all data science tasks. Finally, we show you an end to end project example of a product using all of those components in harmony.
DeepLearning is so powerful that it rapidly transforms the Machine Learning space. But DeepLearning is very resource and data hungry. In this talk we show how neural network training works and how it can be parallelized on large scale GPU clusters using inter-model, intra-model, data and pipelined parallelism. We’ll use TensorFlow 2.0 for demonstration purposes.
Short Abstract: TensorFlow was the hype in 2015. But we discovered many flaws, most of them addressed by Keras and PyTorch. TensorFlow V2 addressed all those and even went beyond what we could imagine in our wildest dreams! This seems to make other frameworks obsolete. In addition, by running natively on Kubernetes using KubeFlow, TensorFlow V2 has the potential to become the de-facto standard for any AI project
Abstract: Towards the End of 2015 Google released TensorFlow, which started out as just another numerical library, but has grown to become a de-facto standard in AI technologies. TensorFlow received a lot of hype as part of its initial release, in no small part because it was released by Google. Despite the hype, there have been complaints on usability as well. Especially, for example, the fact that debugging was only possible after construction of the static execution graph. In addition to that, neural networks needed to be expressed as a set of linear algebra operations which was considered as too low level by many practitioners.
PyTorch and Keras addressed many of the flaws in TensorFlow and gained a lot of ground.
TensorFlow 2.0 successfully addressed those complaints and promises to become the go-to framework for many AI problems.
In this talk we’ll introduce you to the most prominent changes in TensorFlow 2.0 and how you can use these new features successfully in your projects. We’ll cover eager execution, parallelisation strategies, the advantages of the tight high level Keras integration, live neural network training monitoring using TensorBoard, automated hyper parameter optimization, Model serving with TensorFlow Serving, TensorFlow.js and TensorFlow Lite. We’ll finalise with an outlook on TFX - where Google is planning to open source it’s complete AI pipeline and will contrast it with existing de-facto standard frameworks like Apache Spark.
At the end we show how to run TFX Pipelines on Kubernetes using Kubeflow.
What's the takeaway for the audience?
- Learn how TensorFlow Graphs work
- Understand how Eager Execution makes graphs obsolete
- Apply parallelisation strategies to run on clusters
- Use TensorBoard for live neural network debugging
- Execute hyper parameter optimisation
- Integrate TensorFlow Serving, TensorFlow.js and TensorFlow Lite into your projects
- Scale TFX Pipelines on Kubernetes with KubeFlow
Prerequisite knowledge for this presentation?
- Basic machine learning
- Basic python
Apache Spark is the de-facto standard for massive parallel data processing at Enterprise and Cloud Scale - after an introduction in how Apache Spark works let’s have a look on what’s new in Version 2.4 and how the new features are impacting the Big Data and AI ecosystem from an architectural point of view.
TensorFlow is an awesome library. But for the average developer fiddling with linear algebra is far to complicated. In this talk we'll give you a fast track recipe to master DeepLearning challenges using the Keras framework on top of TensorFlow. We'll start with basic image classification, show how you can implement a chat- bot and end with a Cryptocurrency price predictor. At the end of this talk you'll know how Convolutional Neural Networks, Long-Short-Term Memory Networks and Autoencoders work and how you can apply them using Keras and TensorFlow.
We have tried many things. All of those work to a certain extent. Some make us stronger, some make us faster, some make us sick. The ability to think always has been a major challenge. Consciousness is the state or quality of awareness or of being aware of an external object or something within oneself. At least this is the definition on wikipedia. Brute-force state space exploration in chess? Attacking a knowledge base with grammar? Blue Gene and Watson did the job. Artificial neural tissue running In Silico is the latest craze - outperforming human baseline in various cognitive tasks. Let’s extrapolate the latest developments in that space and project. Even conservatively minded folks will notice that things can get out of control. Actually they are already. Or aren’t they? In this talk we’ll cover state of the art AI in robotics and decision making to find out where we are - between zero and singularity. And what steps need to be done to design an AI that humans need. In an open, decentralized, interconnected and unbiased way.
When training a deep neural network, concepts we know as humans are buried into weights between neurons. So when training an image classifier, concepts of dogs and cats are known by the neural network. But these concepts have been in-accessible. GANs are different. GANs can create new data using these hidden concepts. Let's talk about how they work and discuss their applications.
As a Sun Certified Java Programmer and Developer, and after working in 100s of JEE/WebSphere projects for IBM I never could imagine using anything else than Java. Even when planning for the first coursera course - "Fundamentals of Scalable DataScience" - we had tons of discussions whether we should use Scala, R or python for it. Now Python is an obvious choice for data science and we are using it for our "Advanced Machine Learning and Signal Processing" course as well as for the "Applied AI using DeepLearning" course. So in this talk I'll give you a summary on how to use Python and ApacheSpark on top of different machine learning and deep learning frameworks like SparkML, SystemML, Keras, DeepLearning4J to build scalable brains for AI. I'll finalise with a summary on Fabric for DeepLearning and the Docker/Kubernetes based Open IBM Model Asset Exchange.
Summary - This talk explains how machine learning and deep learning can be made accessible to the broader ecosystem by providing consumable services in a micro services infrastructure.
This session explains one of the most important steps in creating a data product. Most DataScience projects nowadays start with a notebook. But in order to be consumable by others model scoring but also model training has to be consumable, even if you don’t have a Ph. D. in math. This talk shows how models implemented in the most common machine learning and deep learning frameworks can be provided as micro service in a scalable and fault tolerant way using container infrastructure on top of GPUs.
NodeRED is one of IBM's recent contribution to OpenSource. We've seen examples of NodeRED processing 1.000.000 msg/s on 64 MB main memory footprint. Learn how you can use NodeRED on the Cloud (any cloud vendor is supported), on your machine, in your data center or even on an Raspberry Pi